Danijela Markovic - Physics for Neuromorphic Computing

Dates
23/10/2024

Danijela Markovic (CNRS - Thales)

Horaires

15h45-17h30

Lieu

Amphi Lhuillier

The brain exhibits remarkable information processing capabilities that far surpass those of digital computers. Inspired by the brain, artificial neural networks (ANNs) have propelled the success of machine learning. ANNs can learn from data without being explicitly programmed, but they require massive datasets for training, suffer from issues like catastrophic forgetting during retraining and consume substantial amounts of energy—sometimes exceeding the power usage of entire states.

Neuromorphic computing draws deeper inspiration from the brain at a physical level, using hardware that mimics the behavior of neurons and exploits physical phenomena such as synchronized oscillations or frequency modulation for learning. Early approaches in neuromorphic computing leveraged unconventional devices like memristors to emulate the brain's distributed architecture, with neurons serving as computational units interconnected by synapse-like memory elements. Recent advances have introduced the use of spintronics and integrated photonics, enabling dynamic processes to enhance learning. Innovations such as encoding input data in injection-locked tones, using phase transition materials for synapses, and employing techniques like equilibrium propagation and coupled learning1 for single-shot training have expanded the possibilities for neuromorphic computing.

More recently, quantum physics has entered the field of neuromorphic computing. The vast Hilbert space of quantum systems allows for the embedding of input data into high-dimensional feature spaces, thus enabling highly compact neural networks. Recent research has shown that quantum entanglement enhances learning and reduces the number of neurons needed for specific tasks2. Furthermore, quantum neural networks were shown to be capable of recognizing quantum input data3, detecting quantum phase transitions4, learning temporal quantum tomography5, and identifying optimal bases for detecting squeezing6. These observations open exciting perspectives for better understanding quantum information processing, brain functioning, and their applications in computing.

References

  1. Stern, M., Hexner, D., Rocks, J. W. & Liu, A. J. Supervised Learning in Physical Networks: From Machine Learning to Learning Machines. Phys. Rev. X 11, 021045 (2021).
  2. Dudas, J. et al. Quantum reservoir compu[ng implementa[on on coherently coupled quantum oscillators. Npj Quantum Inf. 9, 64 (2023).
  3. Ghosh, S., Opala, A., Matuszewski, M., Paterek, T. & Liew, T. C. H. Quantum reservoir processing. Npj Quantum Inf. 5, 35 (2019).
  4. Herrmann, J. et al. Realizing quantum convolu[onal neural networks on a superconduc[ng quantum processor to recognize quantum phases. Nat. Commun. 13, 4144 (2022).
  5. Tran, Q. H. & Nakajima, K. Learning Temporal Quantum Tomography. Phys. Rev. Le:. 127, 260401 (2021).
  6. Khan, S. A., Hu, F., Angelatos, G., Hatridge, M. & Türeci, H. E. A neural processing approach to quantum state discrimina[on. Preprint at hdp://arxiv.org/abs/2409.03748 (2024).